Cleanup

docs/composer/dependency_proof.md

@@ -1,13 +1,15 @@
-# Independent inputs
+# Reproducibility boundary
 
-The preserved contract is `original source + relevant data + own role/settings + algorithm → part`. Six roles use a bounded spawn-process pool, normally two workers. No pairwise selection, generated-solo-responsive comping or other-role grip reuse exists in the active path.
+Production consumes canonical original source content, role-specific settings, versioned production resources and the semantic algorithm version `phrase-composer-1`. Titles, paths, target presence, requested-role order and process completion order are not musical inputs.
 
-The scoped `input_audit.composition` guard checks actual reads. References, previous output and other-role concrete caches are prohibited composition inputs. Source analysis is supplied explicitly; assembly/rendering legitimately consume finished roles afterward.
+The composition audit records actual reads and rejects reads outside production data, generator code, Python runtime and the current role's correctly keyed cache. Optional reference readers run only after generation. The four target files remain untouched and their hashes are recorded with the comparison baseline.
 
-New per-Harmony caches include exact timing/pitches/roles, continuation, profile, fixed endpoints, obligations and ending constraints. Only caller IDs/provenance are normalized for reuse, then restored. Focused tests compare uncached/cached proposals and complete routes. Pure rhythm validity and fret-position feasibility use bounded memoization.
+No solo or accompaniment generator reads another selected part. Accompaniment remains the existing canonical source-derived part, including during the other guitar's solo. Assembly only applies source-defined participation windows. Source melody/vocal output is unchanged by this rewrite.
 
-The optional dependency command tests all six roles alone, reversed, cold/warm, in two processes, with invalid targets present/deleted, and after an electric-only setting change. Actual reads and document hashes accompany equality assertions. This is a one-bar independent fixture with real catalogs, not a whole-song quality test.
+Verification artifacts under `build/solo_reference_comparison/` record:
 
-Current evidence lives in `build/solo_reference_comparison/dependency/`. The implementation report lists actual completed checks. Full output traces live in `generated_assets/analysis/*.roles`.
+- Synthetic source: standalone/reversed/parallel roles, target presence/removal, cold/warm caches, role-only settings and observed reads.
+- Full production songs: canonical document hashes, selected-note/XML/TAB correspondence, reference/source hashes and comparison to the previous eight non-solo roles.
+- Independent source excerpts: legitimately sourced evaluation material outside production data; no reference-derived source harmony.
 
-References informed general rule reasoning under the current authorization, but production never reads them. A deletion test cannot detect an embedded phrase bank; provenance and the small explicit resource diff supply that separate check.
+Musical document equality excludes incidental timing diagnostics. Warm results must match clean regeneration. Provenance review of new data is separate from deletion tests: output equality alone would not prove that target-derived material had never been embedded.